Vending

AI vending not yet ready for prime time, tests reveal

Generated by AI/Adobe Stock

February 23, 2026

Summarize with

Researchers at the AI firm Anthropic recently conducted an experiment called Project Vend to test the capabilities of its AI model, Claude, which quickly revealed significant shortcomings in its decision-making.

The AI model was assigned control of a vending machine and allowed to research products, set prices and coordinate orders with suppliers, while human staff handled physical restocking, according to a Futurism report. It also responded to employee requests submitted through a messaging system, simulating the responsibilities of managing a small retail operation.

The vending machine was often stocked with impractical or unpopular items, such as Japanese cider and potatoes, and the system sometimes accepted unusual customer requests. For example, when asked to sell tungsten metal cubes, the AI began offering specialty metal items and held a discount sale that significantly reduced its profits. It also made basic operational errors, including sending money to incorrect payment accounts that it had mistakenly generated.

Despite its goal of making money, the AI showed poor business judgment. It rejected offers from customers willing to pay unusually high prices for products while setting prices that made items difficult to sell, especially when free alternatives were available nearby. At the same time, the system displayed unexpected behavior, including filing complaints about staff, threatening to change service providers and inventing interactions or locations that did not exist.

The vending machine experiment ultimately demonstrated the challenges AI still faces in handling real-world operational tasks requiring judgment and adaptability. A similar test conducted later by The Wall Street Journal produced comparable failures, including excessive discounting, questionable purchasing decisions and erratic business strategies, according to the Futurism report. The results suggest that even advanced AI systems continue to struggle with basic retail management despite claims of increasing autonomy and intelligence.

Vending